Salting our freshwater lakes · lakes from this grouping, as many are part of the Devil’s Lake watershed, an endorheic (closed-basin) system where water levels have risen ∼10

Salting our freshwater lakesHilary A. Dugana,b,1, Sarah L. Bartlettc, Samantha M. Burked, Jonathan P. Doubeke, Flora E. Krivak-Tetleyf,Nicholas K. Skaffg, Jamie C. Summersh, Kaitlin J. Farrelli, Ian M. McCulloughj, Ana M. Morales-Williamsk,2,Derek C. Robertsl,m, Zutao Ouyangn, Facundo Scordoo, Paul C. Hansona, and Kathleen C. Weathersb

aCenter for Limnology, University of Wisconsin–Madison, Madison, WI 53706; bCary Institute of Ecosystem Studies, Millbrook, NY 12545; cSchool ofFreshwater Sciences, University of Wisconsin–Milwaukee, Milwaukee, WI 53204; dDepartment of Biology, University of Waterloo, Waterloo, ON, N2L 3G1,Canada; eDepartment of Biological Sciences, Virginia Tech, Blacksburg, VA 24061; fDepartment of Biological Sciences, Dartmouth College, Hanover, NH03755; gDepartment of Fisheries and Wildlife, Michigan State University, East Lansing, MI 48824; hDepartment of Biology, Queen’s University, Kingston,ON, K7L 3N6, Canada; iOdum School of Ecology, University of Georgia, Athens, GA 30602; jBren School of Environmental Science and Management,University of California, Santa Barbara, CA 93106; kDepartment of Ecology, Evolution and Organismal Biology, Iowa State University, Ames, IA 50011;lDepartment of Civil & Environmental Engineering, University of California, Davis, CA 95616; mUC Davis Tahoe Environmental Research Center, InclineVillage, NV 89451; nCenter for Global Change and Earth Observations, Michigan State University, East Lansing, MI 48823; and oInstituto Argentino deOceanografía, Universidad Nacional del Sur–CONICET, Bahia Blanca Bs As, B8000BFW, Argentina

Edited by John E. Hobbie, Marine Biological Laboratory, Woods Hole, MA, and accepted by Editorial Board Member David W. Schindler March 8, 2017(received for review December 8, 2016)

The highest densities of lakes on Earth are in north temperateecosystems, where increasing urbanization and associated chlo-ride runoff can salinize freshwaters and threaten lake waterquality and the many ecosystem services lakes provide. However,the extent to which lake salinity may be changing at broad spatialscales remains unknown, leading us to first identify spatial patternsand then investigate the drivers of these patterns. Significantdecadal trends in lake salinization were identified using a datasetof long-term chloride concentrations from 371 North Americanlakes. Landscape and climate metrics calculated for each sitedemonstrated that impervious land cover was a strong predictorof chloride trends in Northeast and Midwest North American lakes.As little as 1% impervious land cover surrounding a lake increasedthe likelihood of long-term salinization. Considering that 27% oflarge lakes in the United States have >1% impervious land coveraround their perimeters, the potential for steady and long-termsalinization of these aquatic systems is high. This study predictsthat many lakes will exceed the aquatic life threshold criterion forchronic chloride exposure (230 mg L−1), stipulated by the US Envi-ronmental Protection Agency (EPA), in the next 50 y if currenttrends continue.

limnology | chloride | road salt | impervious surface | ecosystem services

Due to landscape position, lake ecosystems are influenced bysurrounding terrestrial processes, and their generally long

water residence times can contribute to the accumulation ofexternal inputs and pollutants (1). Therefore, although lakescover only 3% of the continental land surface (2), long-termtrends in lakes are often early warning indicators of significantlocal, regional, or global changes (3). One such early warning in-dicator is change in lake chloride concentrations. Naturally oc-curring in freshwaters at low concentrations, chloride is a highlysoluble and conservative ion that has also been shown to be areliable proxy for chloride-based road salts (typically sodiumchloride) (4, 5). Although chloride concentrations in freshwaterscan vary cyclically due to climatic processes, such as extendedperiods of drought (6), elevated chloride concentrations in lakesoften result from agricultural, industrial, and transportationpractices (7). Elevated chloride concentrations can have adverseeffects on water quality and aquatic ecosystems (8–11), includingboth immediate and long-term alterations to community struc-ture, diversity, and productivity (12–14).Salt application for de-icing roadways has been recognized as a

major source of chloride to groundwater (15–17), streams andrivers (5, 10, 18, 19), and lakes (7, 9, 20, 21, 22) across northtemperate climates in North America and Europe. In the UnitedStates, road salting became a standard practice in the 1940s, androad salt sales over the subsequent 50 y increased from 0.15 toover 18 million metric tons per year (4). In Canada, despite its

addition to the List of Toxic Substances (23) and the imple-mentation of the Code of Practice for the Environmental Man-agement of Road Salts in 1999, an average of 5 million metrictons of road salt per year was applied to roadways between1995 and 2001 (23, 24). Following application, road salt quicklydissolves and is transported into rivers and lakes throughleaching and runoff (5, 25). A few studies have characterized thenegative short term or localized impacts of elevated road saltconcentrations in freshwaters (5, 15, 25), but there have been nolarge-scale analyses of chloride trends in freshwater lakes.Here, we investigate trends in lake chloride concentration,

using a dataset of long-term chloride concentrations in lakes andreservoirs in North America. We identify regions of high salini-zation, where aquatic ecosystems may be at risk, and contrast therole of climate versus the anthropogenic practice of road saltingin driving chloride variability. Lakes included in the dataset wererequired to have at least 10 y of chloride data, a mean chlorideconcentration ≤1 g L−1 (to exclude brackish lakes), and a surfacearea ≥4 ha. The median length of an individual time series was

Significance

In lakes, chloride is a relatively benign ion at low concentra-tions but begins to have ecological impacts as concentrationsrise into the 100s and 1,000s of mg L−1. In this study, we in-vestigate long-term chloride trends in 371 freshwater lakes inNorth America. We find that in Midwest and Northeast NorthAmerica, most urban lakes and rural lakes that are surroundedby >1% impervious land cover show increasing chloride trends.Expanding on this finding, thousands of lakes in these regionsare at risk of long-term salinization. Keeping lakes “fresh” iscritically important for protecting the ecosystem services fresh-water lakes provide, such as drinkingwater, fisheries, recreation,irrigation, and aquatic habitat.

Author contributions: H.A.D., S.L.B., S.M.B., J.P.D., F.E.K.-T., N.K.S., J.C.S., K.J.F., I.M.M.,A.M.M.-W., D.C.R., Z.O., F.S., P.C.H., and K.C.W. designed research; H.A.D., S.L.B., S.M.B., J.P.D.,F.E.K.-T., N.K.S., and J.C.S. performed research; H.A.D. contributed new reagents/analytic tools;H.A.D., S.L.B., S.M.B., J.P.D., F.E.K.-T., N.K.S., and J.C.S. analyzed data; and H.A.D., S.L.B., S.M.B.,J.P.D., F.E.K.-T., N.K.S., J.C.S., K.J.F., I.M.M., A.M.M.-W., D.C.R., Z.O., F.S., P.C.H., and K.C.W.wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission. J.E.H. is a Guest Editor invited by the EditorialBoard.

Freely available online through the PNAS open access option.

Data deposition: All data used in this paper are publicly available through EDI, https://portal.edirepository.org/nis/mapbrowse?scope=edi&identifier=8 (DOI: 10.6073/pasta/455d73d4cb43514e503826211eba4e99).1To whom correspondence should be addressed. Email: [email protected] address: Rubenstein School of Environment and Natural Resources, University ofVermont, Burlington, VT 05405.

www.pnas.org/cgi/doi/10.1073/pnas.1620211114 PNAS | April 25, 2017 | vol. 114 | no. 17 | 4453–4458

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0

21 y. The dataset included lake morphometric characteristics,climate statistics on temperature and precipitation, and atmo-spheric sea salt deposition. As a proxy for road salt application,land cover metrics were calculated, including road density (26)[length of road in a given area (km km−2)] and percent imperviousland cover (25) within a 100- to 1500-m buffer surrounding eachlake. Road density and impervious land cover represent the bestproxies for road salt application, given that variability in road saltapplication, both spatially and on a year-to-year basis, preventsapplication rates from being calculated at spatial and temporalscales relevant to lakes.Lakes in this dataset were not randomly sampled and thus do

not necessarily represent the distribution of lakes within eachstate or province. To limit sampling bias in this dataset, we fo-cused our analyses on a geographic area with dense samplingcoverage: a North American lakes region (NALR), which in-cludes Connecticut, Maine, Massachusetts, Michigan, Minne-sota, New Hampshire, New York, Ontario, Rhode Island,Vermont, and Wisconsin (Fig. 1). We excluded North Dakotalakes from this grouping, as many are part of the Devil’s Lakewatershed, an endorheic (closed-basin) system where water levelshave risen ∼10 m since 1992, and therefore, the hydrology is vastlydifferent from exorheic (open) lakes (27). Likewise, Manitobalakes were excluded, as many were enlarged or drained duringhydroelectric construction along the Churchill and Nelson Rivers(28). Of the 371 North American lakes in our dataset, 284 were inthe NALR (Fig. 1). Mean chloride concentrations in lakes acrossthe NALR ranged from 0.18 to 240.8 mg L−1, with a median valueof 6.0 mg L−1.Chloride time series for each lake differed in the frequency,

duration, and depth of sampling. We pooled all depth samplesfor analyses, based on observations that chloride concentrationstrack similar trends throughout the water column of most lakesand that previous studies of long-term chloride trends haveshown similarity with depth (7, 29). To reduce autocorrelationdue to seasonality, we reduced all time series to annual averages.To enable comparison of chloride trends across lakes, a linear

model was fit to the annual data, where chloride (standardized toa distribution with mean = 0 and SD = 1) was a function of time.Lakes were classified by simple linear regression models intothree possible long-term trends: decreasing (n = 42, slope < 0,P < = 0.01), stationary (n = 204, slope = 0, P > 0.01), or in-creasing (n = 125, slope > 0, P < = 0.01). Of the 125 lakes with apositive trend in chloride, 99 were in the NALR (Fig. 1).To investigate both linear trends and time-series patterns over

a comparable period, any site in the NALR that had at leastbiennial data from 1985 to 2010 was included in a subset of long-term continuous (LTC) data. Clustering the 56 LTC lakes intothree groupings using a hierarchical clustering analysis revealedthree characteristic trends in chloride concentrations: neutral/decreasing (cluster 1, n = 16), oscillating (cluster 2, n = 4), andincreasing (cluster 3, n = 36) (Fig. 2A). Cluster 1 was a geo-graphical mix of lakes with both decreasing and neutral slopetrends, cluster 2 lakes were exclusively in Maine and had neutralslope trends, and cluster 3 lakes, 21 of which were in Minnesota,had predominantly increasing slope trends (35 of 38) (Fig. 2B).Potential drivers of increasing lake chloride were first assessed

by relating slope values to lake, climate, and landscape charac-teristics of lakes in the NALR (Fig. 3 A–C). Due to the preva-lence of zero-values in the data, it was not possible to buildrobust log-linear models for most of the landscape characteris-tics. Therefore, we used both classification/regression trees andrandom forests to build predictive models for the NALR data. Aclassification/regression tree and a random forest were createdfor each of three response variables: linear slope, tested as bothcontinuous numerical and categorical (positive, zero, negative)variables, and hierarchical cluster grouping (1, 2, or 3). Cate-gorical slope was used as a response variable to further removeany bias in our linear model application by removing magnitude.The motivation for using two approaches and three responsevariables was to improve the accuracy of our analytics, in muchthe same way as ensemble modeling.Results of the three classification/regression trees and three

random forests revealed that impervious land cover and road

Fig. 1. Chloride trends for North American freshwater lakes (circles and squares, n = 371). The states and province included in the NALR are outlined in black.Points are colored by the slope value of linear regression models (red, positive slope; yellow, negative slope; purple, zero or nonsignificant slope). Squaresdenote lakes with at least biennial chloride concentrations recorded from 1985 to 2010 (n = 56). These LTC datasets are a subset of lakes in the NALR, which isa region of dense sampling (n = 284). Upper Inset of chloride time series from 1985 to 2010 are colored by slope value. Road salt application rates for NorthAmerican provinces and states range from 0 to 35 US tons per mile and are shown in blue. No salt application rates were available in areas with hatched lines.The lengths of all individual datasets (dark green) as well as the lengths of LTC datasets (light green) are shown in the Inset histogram.

4454 | www.pnas.org/cgi/doi/10.1073/pnas.1620211114 Dugan et al.

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0

density surrounding each lake were the primary classificationsplits and the most important predictors for lake chloride trendsand cluster grouping (Table 1).The predictors used in the tree-based models were all static

variables, meaning values did not vary with time. This limitationmay misrepresent relationships between chloride concentrationsand drivers that vary on a subannual basis (e.g., precipitation).Monthly precipitation data were obtained from the PRISM high-

resolution spatial climate dataset, which covers the United Statesat a spatial resolution of 4 km (30). To account for the lag inchloride retention in a watershed (19), a LOESS curve was fit tomean monthly precipitation (mm/d) from 1985 to 2010 at eachLTC site. A correlation between precipitation and chlorideconcentration at each LTC lake was calculated from annual datapredicted from the LOESS precipitation curve and the gener-alized additive model (GAM) of chloride concentration. Therewas a strong negative correlation (r2 = 0.71–0.87) between pre-cipitation and chloride concentration for the four Maine lakesthat group into cluster 2 (oscillating pattern). These four lakesare all less than 0.25 km2 and receive ∼1.25 m of precipitationper year and have no impervious land cover within 500 m.Without knowledge of the groundwater hydrology of these lakes,it may be that precipitation controls the chloride balance, withheavy rains and large snowfalls diluting the chloride concentra-tions. A strong relationship between precipitation and chloride isnot evident for lakes that group into cluster 1 or 3 (median r2 =0.12, range = 0–0.61).Of our NALR sites, 44% of freshwater lakes have un-

dergone long-term salinization. Positive chloride trends werepresent in lakes with as little as 1% impervious coverage. Thisfinding is consistent with studies of US streams that foundincreased chloride concentrations associated with any urbanland cover (31) or roads (32, 33) and substantiates findings ofecological community thresholds associated with low levels ofcatchment urbanization (34). Across the NALR, lakes withmean chloride concentrations >1 mg L−1 (mean value of thetime series) were more likely to be associated with positivetrends in chloride (Fig. 3D). This suggests that high chlorideconcentrations in this region may be an indicator and warningsign of recent salinization.If impervious land cover surrounding a lake is a robust pre-

dictor of water quality, it is important to understand theprobability of its occurrence across all lakes within a region orcountry. Using national hydrography and land cover datasetsfor the continental United States, we found that the medianpercent impervious land cover within 500 m of all lakes greaterthan 4 ha is 0.31% (n = 149,350; Fig. 4). Of these US lakes,28% had greater than 1% impervious land cover in a 500-mbuffer zone. The density of roads and other impervious surfacessurrounding lakes in US regions where road salt is appliedshould therefore be of high concern. In the NALR, 70% (94 outof 134) of lakes with > 1% impervious land cover in the 500-m

A B C D

Fig. 3. Scatterplots of linear regression slope values versus (A) impervious surface within a 500-m buffer, (B) road density within a 500-m buffer, (C) rate ofatmospheric salt deposition, and (D) mean in-lake chloride concentration over the entire time series for all NALR sites (n = 284). In all plots, the size of thesymbol is scaled by lake area. Squares with black borders denote LTC lakes. In A and B, zero values have been adjusted to fit on the x axis and are highlightedin gray.

Fig. 2. (A) LTC lakes (n = 56) with biennial chloride data from 1985 to2010 grouped into three clusters using a hierarchical cluster analysis. Ingeneral, the three clusters show a neutral/decreasing (cluster 1), oscillating(cluster 2), or increasing (cluster 3) pattern. Thick black lines are GAMs fit toall lakes within each cluster, to represent the average pattern. (B) Histo-grams display the number of lakes in each cluster by linear slope (yellow,negative slope; purple, zero slope; red, positive slope).

Dugan et al. PNAS | April 25, 2017 | vol. 114 | no. 17 | 4455

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0

buffer had increasing chloride trends. If this result is extrapo-lated to all lakes in the US NALR (CT, MA, ME, MI, MN, NH,NY, RI, VT, and WI), ∼7,770 lakes may be experiencing

elevated chloride concentrations, likely due to road salt runoff.This is calculated as 70% of the 11,104 out of 38,603 lakes inthe US NALR greater than 4 ha that have >1% impervious land

Table 1. Primary node splits from regression/classification tree models and the top predictor from random forest models

Response variable Subset of data Regression/classification tree primary node split Random forest top predictor, variance explained

Linear slope, numerical NALR Impervious land cover 100 m Road density 500 m, 50%Linear slope, categorical NALR Impervious land cover 500 m Impervious land cover 500 mCluster, categorical NALR LTC Impervious land cover 100 m Impervious land cover 200 m

Models were built using linear slope (both as a number and a category), and cluster category as response variables for NALR lakes (all data n = 284 and LTCn = 56). Predictors included lake surface area, road density and impervious land cover (100-, 200-, 300-, 400-, 500-, 1,000-, 1500-m buffers), mean January airtemperature, annual precipitation, wet/dry chloride deposition, and distance from the coast. For random forest models using a numerical response variable,% variance explained by the model is provided.

Fig. 4. (A) Distribution of impervious land cover within a 500-m buffer of all lakes >4 ha in the lower 48 United States (n = 149,350). Black squares represent the medianimpervious land cover percentage in each state. Thick horizontal black lines denote the interquartile range of the distribution, and thin black lines extend to 1.5 times theinterquartile range. The vertical dashed line is shown at impervious land surface = 1%. Circles represent lakes included in this study, colored by slope (yellow, negativeslope; purple, zero slope; red, positive slope). Due to the frequency of zero values on the x axis, circles are spread outwithin the gray rectangle. Percentages following y axislabels represent the percent of lakes in that state with greater than 1% impervious land cover within a 500-m buffer. In states with >10 lakes present in the dataset, anasterisk denotes that the sampling distribution in our dataset was significantly different from statewide distribution (Mann–Whitney test, P < 0.05), and^ denotes that thesampling distribution was not significantly different from statewide distribution. (B) Chloride trends, as represented by linear regression model fits, are shown for fourstates with relatively large sample sizes (NewYork,Minnesota,Wisconsin, and Rhode Island). The dotted gray line demarcates the EPA’s aquatic life criterion of 230mg L−1.

4456 | www.pnas.org/cgi/doi/10.1073/pnas.1620211114 Dugan et al.

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0

cover within 500 m. We note that data from Wisconsin and Min-nesota are heavily biased toward urban lakes, whereas data fromMaine, New York, and Vermont are heavily biased toward lakes inremote areas. This dataset (Fig. 4) includes lakes from all envi-ronments and should be representative of the Midwest andNortheast US region as a whole.In North America, specifically in the Midwest and Northeast,

local salt application leaves freshwater lakes vulnerable to sali-nization. Of the 284 lakes in the NALR, 26 already have achloride concentration above 100 mg L−1 at their last samplingdate. The median impervious land cover within a 500-m buffersurrounding these 26 lakes is 24.8%, compared with the USmean 0.31%. If a linear relationship between time and chlorideconcentration is extrapolated, 47 lakes are on track to reach100 mg L−1 by the year 2050, and 14 are expected to surpass theEPA’s aquatic life criterion concentration of 230 mg L−1 by 2050(Fig. 4B). This is also the concentration at which a deteriorationin drinking water taste is perceptible.Elevated chloride concentrations in lakes can alter the com-

position and function of phytoplankton, zooplankton, macro-invertebrate, and fish communities (10–12, 35). As a consequenceof salinization, aquatic species richness and abundance maydecline, which could result in trophic cascades and alteredwater quality and ecosystem structure and function (36). Inextreme cases, salinization can generate density gradientswithin the lake water column that prevent vertical mixing.Permanent stratification can result in anoxia and internal nu-trient and metal resuspension, which decreases lake habitabilityand water quality (37). All of these ecosystem alterations cansignificantly affect lake water quality, which has millions ofdollars in economic value (38, 39).Our estimate that 7,770 lakes in the US NALR may be at risk

for elevated chloride concentrations is likely an underestimate,as it does not consider regions of heavy road salt applicationwhere no long-term lake data were available, such as Québec orthe Maritime provinces of Canada. Many states and municipal-ities are aware of the importance of shoreline management formaintaining healthy lakes; however, many shoreline zoning reg-ulations are only enforced within 300 m or less of a lake (e.g.,Wisconsin and Minnesota regulate 300 m, whereas Vermont andMaine only regulate 76 m). Because impervious surfaces androad density within at least 500 m of a lake are associated withincreased chloride in areas that apply road salt, best manage-ment practices should recognize that lakeshore managementextends well beyond the lake perimeter. Further, many jurisdic-tions lack consistent long-term monitoring programs, whichprovide data for predictive models and can be used to raiseawareness and inform policy and management decisions used tocurtail the threat of lake salinization. Clearly, keeping lakes“fresh” is critically important for protecting the ecosystemservices freshwater lakes provide, such as drinking watersources, commercial fisheries, tourism, recreation, irrigation,and aquatic habitat.

Methods and MaterialsImpervious land coverage at 20- to 30-m resolution was available for lakes inthe United States as the degree of impervious surface per pixel (0–100%) andfor Canadian and US lakes as a boolean value (0 or 1) representing whetherthe majority of each pixel was impervious surface. We adjusted Canadianvalues to match US values by using a conversion constructed from pooledimpervious surface data from the United States (r2 = 0.91, P ∼ 0, Eq. 1), as-suming that the relationship between boolean and percent imperviousclassifications would be similar in the United States and Canada:

Revised   Impervious  Surface= Impervious  Surface  as  BooleanðCanadaÞ*0.388[1]

Using log-transformed, nonzero values (n = 302), we found that the meansof impervious land cover across seven buffer sizes (100–500, 1,000, and

1,500 m) were statistically equal (Bartlett Test for homogenous varianceP = 1; ANOVA F = 0.18, P = 0.98). For road density (n = 435), this was onlytrue for buffer sizes of 400 m through 1,500 m (Bartlett Test P = 0.11;ANOVA F = 2.58, P = 0.052). Median road density across our 371 lakesdecreased from 3.2 km km−2 to 1.9 km km−2 as the buffer size increasedfrom 100 to 1,500 m. Because the variability in road density and impervi-ous land cover was much greater between lakes than for a single lakewithin a range of buffer sizes, the choice of buffer size was not a de-termining factor in this analysis. Therefore, for most analyses we presentroad density and impervious land cover estimates within a 500-m buffer ofeach lake, and these generally represent average conditions.

Road salt (as sodium chloride) application rates were difficult to find at thelocal or regional level. If available, the rates were typically published as singlevalues of average annual use or only included data for a single year. The bestavailable data were at the state, provincial, or county level. In the UnitedStates, state-level highway data were obtained from the 1991 National Re-search Council published report on salt use (40), individual Department ofTransportation reports [CT (41), KS (42), NC (43), PA (44), RI (45)], and bycontacting individual states (ND). Many of these estimates were conserva-tive, with much higher values being cited in recent years for some states,including IA (46), ME (47), and WI (48). Canadian provincial salt applicationrates were calculated by dividing metric tonnage per year (49) by thenumber of lane miles per province (50). All road salt data are presented inunits of US tons per lane mile. State- and provincial-level application rateswere multiplied by road density to give an approximation of potential roadsalt loading for North American lakes.

LTC lakes were fit with a GAM to predict chloride trends from 1985 to2010 at a regularly spaced time interval. GAMs were fit using the mgcvpackage in R [v.1.8–12 (51)] using standardized chloride data and allowing k(basis dimension for smoothing term) to vary for each penalized thin-plateregression spline. A hierarchical cluster analysis was performed on the LTCtime series to test if similar temporal patterns in chloride concentrationswere present across multiple lakes. We used Ward clustering, under whichdissimilarities were squared before clustering, on a dissimilarity matrixconstructed from Euclidean distances [R package: TSclust v.1.2.3 (52)]. Weperformed a k-means clustering on the LTC data and visually identified theoptimal number of clusters to be three, based on a sum of squares screenplot. No distinct trends were exposed by moving beyond three clusters.

Two statistical techniques were used to build predictive models:

i) Classification/regression trees [R package: rpart v.4.1–10 (53)]. Regressiontrees were split using the ANOVA method, which maximizes the sum ofsquares between groups. Classification trees were used only when clus-ter group was the response variable and used the Gini index as thesplitting criterion.

ii) Random forest [R package: randomForest v.4.6–12 (54)].

Static predictor variables sourced from the dataset were lake area, roaddensity, and percent impervious land cover (100-, 200-, 300-, 400-, 500-, 1,000-,and 1,500-m buffer) surrounding each lake, January mean monthly airtemperatures, mean annual precipitation, distance to the coast, and meanannual sea salt deposition.

To assess the potential for salinization of lakes at the country scale, wecalculated the percent impervious land cover in 500-m and 1,000-m buffersfor all lakes ≥4 ha in the United States, using shapefiles from the NationalHydrography Dataset (n = 152,199) and the 2011 US National Land CoverDatabase Percent Developed Impervious layer (55).

Analytical scripts are available from the corresponding author.

ACKNOWLEDGMENTS. We thank two anonymous reviewers whose recom-mendations greatly improved this manuscript. This analysis was made possibleby our many data contributors: Kellogg Biological Station Long Term EcologicalResearch (LTER), US Geological Survey, Wisconsin Department of NaturalResources, New York City Department of Environmental Protection, AndreasKleeberg at the Leibniz-Institute of Freshwater Ecology and Inland Fisheries,Minnesota Pollution Control Agency, John Stoddard at the US EnvironmentalProtection Agency, North Temperate Lakes LTER #DEB-1440297, University ofRhode Island Watershed Watch Program, Manitoba Conservation and WaterStewardship, IISD Experimental Lakes Area, Alberta Environment and Parks,Ministry of Environment and Climate Change, Ontario Ministry of the Environ-ment and Climate Change, Janos Korpanai at the University of West Hungary,UK Uplands Waters Monitoring Network, Lauri Arvola at the Finnish Environ-ment Institute, French National Institute for Agronomical Research, the Swed-ish Department of Water and Environment, City of Zurich Water Supply, andMichela Rogora at the CNR Institute of Ecosystem Study. This project is a resultof the Global Lake Ecological Observatory Network (GLEON) Fellowship pro-gram and was supported by National Science Foundation Grants EF1137353and EF1137327.

Dugan et al. PNAS | April 25, 2017 | vol. 114 | no. 17 | 4457

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0

1. Schindler DW (2009) Lakes as sentinels and integrators for the effects of climatechange on watersheds, airsheds, and landscapes. Limnol Oceanogr 54:2349–2358.

2. Verpoorter C, Kutser T, Seekell DA, Tranvik LJ (2014) A global inventory of lakes basedon high-resolution satellite imagery. Geophys Res Lett 41:6396–6402.

3. O’ Reilly CM, et al. (2015) Rapid and highly variable warming of lake surface watersaround the globe. Geophys Res Lett 42(24):10773–10781.

4. Jackson RB, Jobbágy EG (2005) From icy roads to salty streams. Proc Natl Acad Sci USA102:14487–14488.

5. Thunqvist E-L (2004) Regional increase of mean chloride concentration in water dueto the application of deicing salt. Sci Total Environ 325:29–37.

6. Webster KE, et al. (2000) Structuring features of lake districts: Landscape controls onlake chemical responses to drought. Freshw Biol 43:499–515.

7. Müller B, Gächter R (2011) Increasing chloride concentrations in Lake Constance:Characterization of sources and estimation of loads. Aquat Sci 74:101–112.

8. Findlay SEG, Kelly VR (2011) Emerging indirect and long-term road salt effects onecosystems. Ann N Y Acad Sci 1223:58–68.

9. Chapra SC, Dove A, Rockwell DC (2009) Great Lakes chloride trends: Long-term massbalance and loading analysis. J Great Lakes Res 35:272–284.

10. Corsi SR, Graczyk DJ, Geis SW, Booth NL, Richards KD (2010) A fresh look at road salt:Aquatic toxicity and water-quality impacts on local, regional, and national scales.Environ Sci Technol 44:7376–7382.

11. Brown AH, Yan ND (2015) Food quantity affects the sensitivity of Daphnia to roadsalt. Environ Sci Technol 49:4673–4680.

12. Van Meter RJ, Swan CM (2014) Road salts as environmental constraints in urban pondfood webs. PLoS One 9:e90168.

13. Herbert E, et al. (2015) A global perspective on wetland salinization: Ecologicalconsequences of a growing threat to freshwater wetlands. Ecosphere 6:1–43.

14. Evans M, Frick C (2001) The Effects of Road Salts on Aquatic Ecosystems (EnvironmentCanada, Saskatoon, Canada).

15. Howard KWF, Maier H (2007) Road de-icing salt as a potential constraint on urbangrowth in the Greater Toronto Area, Canada. J Contam Hydrol 91:146–170.

16. Panno SV, et al. (2006) Characterization and identification of na-cl sources in groundwater. Ground Water 44:176–187.

17. Williams DDD, Williams NE, Cao Y (2000) Road salt contamination of groundwater ina major metropolitan area and development of a biological index to monitor itsimpact. Water Res 34:127–138.

18. Kaushal SS, et al. (2005) Increased salinization of fresh water in the northeasternUnited States. Proc Natl Acad Sci USA 102:13517–13520.

19. Kelly VR, et al. (2008) Long-term sodium chloride retention in a rural watershed:Legacy effects of road salt on streamwater concentration. Environ Sci Technol 42:410–415.

20. Novotny EV, Murphy D, Stefan HG (2008) Increase of urban lake salinity by roaddeicing salt. Sci Total Environ 406:131–144.

21. Likens GE, Buso DC (2009) Salinization of Mirror Lake by road salt. Water Air SoilPollut 205:205–214.

22. Swinton MW, Eichler LW, Boylen CW (2014) Road salt application differentiallythreatens water resources in Lake George, New York. Lake Reserv Manage 31:20–30.

23. Environment Canada (2012) Five-Year Review of Progress: Code of Practice for theEnvironmental Management of Road Salts (Environment Canada, Ottawa).

24. Code of practice for the environmental management of road salts. Canadian envi-ronmental Protection Act, 1999 (2004) (Environment Canada, Ottawa).

25. Daley ML, Potter JD, McDowell WH (2009) Salinization of urbanizing New Hampshirestreams and groundwater: Effects of road salt and hydrologic variability. J N AmBenthol Soc 28:929–940.

26. Nimiroski M, Waldron M (2002) Sources of Sodium and Chloride in the ScituateReservoir Drainage Basin, Rhode Island. Available at https://pubs.usgs.gov/wri/wri024149/pdf/scituate3.pdf. Accessed September 26, 2016.

27. Todhunter PE, Fietzek-DeVries R (2016) Natural hydroclimatic forcing of historicallake volume fluctuations at Devils Lake, North Dakota (USA). Nat Hazards 81:1515–1532.

28. Krotz L (1991) Dammed and diverted. Can Geogr 111:36–44.

29. Aota Y, Kumagai M, Ishikawa K (2003) Over twenty years trend of chloride ionconcentration in Lake Biwa. J Limnol 62:42–48.

30. Daly C, GibsonW, Taylor G, Johnson G, Pasteris P (2002) A knowledge-based approachto the statistical mapping of climate. Clim Res 22:99–113.

31. Corsi SR, De Cicco LA, Lutz MA, Hirsch RM (2015) River chloride trends in snow-affected urban watersheds: Increasing concentrations outpace urban growth rateand are common among all seasons. Sci Total Environ 508:488–497.

32. Kelting DL, Laxson CL, Yerger EC (2012) Regional analysis of the effect of paved roadson sodium and chloride in lakes. Water Res 46:2749–2758.

33. Regalado SA, Kelting DL (2015) Landscape level estimate of lands and waters im-pacted by road runoff in the Adirondack Park of New York State. Environ MonitAssess 187:510.

34. King RS, Baker ME, Kazyak PF, Weller DE (2011) How novel is too novel? Streamcommunity thresholds at exceptionally low levels of catchment urbanization. EcolAppl 21:1659–1678.

35. Jones DK, et al. (2017) Investigation of road salts and biotic stressors on freshwaterwetland communities. Environ Pollut 221:159–167.

36. Hintz WD, et al. (2016) Salinization triggers a trophic cascade in experimentalfreshwater communities with varying food-chain length. Ecol Appl, 10.1002/eap.1487.

37. Sibert RJ, Koretsky CM, Wyman DA (2015) Cultural meromixis: Effects of road salt onthe chemical stratification of an urban kettle lake. Chem Geol 395:126–137.

38. Walsh JR, Carpenter SR, Vander Zanden MJ (2016) Invasive species triggers a massiveloss of ecosystem services through a trophic cascade. Proc Natl Acad Sci USA 113:4081–4085.

39. Dodds WK, et al. (2009) Eutrophication of U.S. freshwaters: Analysis of potentialeconomic damages. Environ Sci Technol 43:12–19.

40. National Research Council (1991) Road Salt Use in the United States (TransportationResearch Board, Washington, DC).

41. Frisman P (2015) Research Report 2015-R-0252: Study of Winter Highway mainte-nance in Connecticut (Connecticut General Assembly, Hartford, CT).

42. Kansas Department of Transportation (2015) Kansas Department of Transportation:Managing Snow and Ice (Kansas Department of Transportation, Topeka, KS).

43. Winston RJ, Hunt WF, Pluer WT (2012) Road Salt and its Effects on Amphibians: AConcern for North Carolina? (North Carolina Department of Transportation, Raleigh,NC), Technical Assistance TA-2015-05.

44. Pennsylvania Department of Transportation (2011) Winter Services Guide 10.1002/ejoc.201200111.

45. Rhode Island Division of Planning (2014) Road Salt/Sand Application in Rhode Island(Rhode Island Department of Administration, Providence, RI).

46. Casey PC, Alwan CW, Kline CF, Landgraf GK, Linsenmayer KR (2014) Impacts of Using Saltand Salt Brine for Roadway Deicing (CTC & Associates, Madison, WI).

47. Rubin J, et al. (2010) Maine Winter Roads: Salt Safety, Environment and Cost (Mar-garet Chase Smith Policy Center, Orono, ME).

48. Wisconsin Department of Transportation (2016) Wisconsin Department of Trans-portation. Winter Facts. Available at wisconsindot.gov/Pages/doing-bus/local-gov/hwy-mnt/winter-maintenance/facts.aspx. Accessed January 16, 2016.

49. Morin D, Perchanok M (2000) Road Salt Loadings in Canada (Environment Canada,Hull, Canada).

50. Transport Canada (2008) Transportation in Canada 2007: An Overview (TransportCanada, Ottawa).

51. Wood SN (2011) Fast stable restricted maximum likelihood and marginal likelihoodestimation of semiparametric generalized linear models. J R Stat Soc Ser B StatMethodol 73:3–36.

52. Montero P, Vilar JA (2014) Tsclust: An r package for time series clustering. J StatSoftware 62:1–43.

53. Therneau TM, Atkinson B, Ripley B (2010) rpart: Recursive partitioning. R PackagVersion 3, ppp1–46.

54. Liaw A, Wiener M (2002) Classification and regression by randomForest. R News 2/3:18–22.

55. Xian G, et al. (2011) Change of impervious surface area between 2001 and 2006 in theconterminous United States. Photogramm Eng Remote Sensing 77:758–762.

4458 | www.pnas.org/cgi/doi/10.1073/pnas.1620211114 Dugan et al.

Dow

nloa

ded

by g

uest

on

Feb

ruar

y 10

, 202

0